Understanding human-level knowledge is fundamental to advancing artificial intelligence, particularly through the incorporation of commonsense knowledge, which includes essential facts about events, beliefs, and desires. This thesis explores the representation and application of commonsense knowledge across various levels—word, phrase, and sentence—utilizing structured knowledge graphs and generative language models. We investigate the evolution of knowledge graphs from lexical resources like WordNet to commonsense knowledge graphs such as ConceptNet and ATOMIC.

Key contributions include addressing training bias in word-level semantic representation using a Z-reweighting strategy, developing a universal sense representation to mitigate data scarcity in multilingual contexts, and propos...[ Read more ]

Understanding human-level knowledge is fundamental to advancing artificial intelligence, particularly through the incorporation of commonsense knowledge, which includes essential facts about events, beliefs, and desires. This thesis explores the representation and application of commonsense knowledge across various levels—word, phrase, and sentence—utilizing structured knowledge graphs and generative language models. We investigate the evolution of knowledge graphs from lexical resources like WordNet to commonsense knowledge graphs such as ConceptNet and ATOMIC.

Key contributions include addressing training bias in word-level semantic representation using a Z-reweighting strategy, developing a universal sense representation to mitigate data scarcity in multilingual contexts, and proposing a multi-alternative contrastive learning framework (MICO) for effective phrase-level representations. Additionally, we examine efficient knowledge reasoning with knowledge sources of both knowledge graphs and language models, through a novel grounding-pruning-reasoning pipeline, enhancing the efficiency of reasoning tasks.

Ultimately, this work not only advances commonsense knowledge understanding and its application in various downstream tasks, including word sense disambiguation, knowledge retrieval, and question answering, but also opens avenues for exploring the application of embodied AI, which can leverage this enriched knowledge for more intuitive and intelligent interactions in real-world settings.